Product
A simulation and decision engine for commercial decisions.
Goal and limits in; your candidate decisions run as versions in a market model tuned to your sales; the plans worth taking come out with a probable range.
- Goal
- Grow contribution margin this month
- Limits
- Discount at most 25%, stock on hand
Market model
Versions
- 15% off, one week
- 20% off, one week
- 20% off, two weeks
Decision15% off, one week
Contribution margin after commission, shipping and unit cost
Illustrative. Numbers come from your run.
What it is
One engine, four parts.
One simulated market, tuned to your products. Pricing, promotions and goal-to-decision run on it.
Goal and limits in
Market model
Shopper response to price and percent-off campaigns, from published research, tuned by your sales history.
Versions
Every decision you are weighing runs in the same market, so the gap between them is the decision, not luck.
Economics
Marketplace commission and shipping per version; compared in contribution margin, not units. Returns and ad cost not modelled.
Decision engine
Goal and limits in, a search across plans, the best few out. Estimates shrunk: the best-looking plan is always overstated.
Decision out
Who it is for
People who carry one decision themselves.
Someone who picks a number this week and lives with it next month.
Marketplace sellers
LiveYou set the price, build the campaign calendar and carry the margin.
Brand pricing and category managers
LiveYou defend a price move to finance with more than last year's volumes.
Restaurant chains
RoadmapMenu prices and combos across branches. Roadmap, not sold as live.
Why this product
The two cheap ways to answer a pricing question both fail.
A language model playing your shoppers is fast and unreliable; a live test is reliable, expensive and moves the customers. Jonbar sits in between: a market model tuned to your sales, versions run before the decision, every forecast locked.
- 40.8/100best language models' score at simulating human behaviour, 20 datasets.SimBench, ICLR 2026
- r = 0.20average correlation between people and their language-model stand-ins.Peng et al., 2025
- 20.3%fall in existing customers' future revenue after a deep discount, field experiment.Anderson & Simester, Marketing Science
- 3 in 4promotions that do not break even, across 39 million US events.Nielsen, via FoodNavigator
Status today
Live, and roadmap.
Post-campaign dip, cannibalisation, stockpiling and competitor match: in v0, off until calibrated.
Live today
- Goal to decision: price, discount, duration search
- Promotion rehearsal: depth, length, start week
- Price steps with demand response and range
- Commission and shipping, scored in contribution margin
- Forecast locked per run, error kept
Roadmap
- Learned shopper population
- Behaviour model trained on purchase sequences
- Live world signals: calendar, prices, platform rules
- Bundles and launches
- Restaurant chains
Details: how we build it, how we check it, your data
How we build it
- 01
The market is modelled, not prompted.
Shopper behaviour, price-to-demand links and platform economics are written as models we can inspect. Language models only explain; they never decide.
- 02
Your history tunes it.
Price response and promotion lift are fitted to your products, with published priors where your history is thin.
- 03
Every forecast is locked before the outcome.
It is compared with actual sales after you act.
How we will know it works
Criteria, not results; no published numbers yet.
Direction on blind tests
MeasuringOn a past campaign we have not seen, does the engine get the direction right.
Size of the error
MeasuringHow far the locked forecast lands from actual units and revenue, against a simple rule.
Decisions changed
MeasuringHow often a rehearsal changes the decision the team was about to make.
Pilots that continue
MeasuringHow many pilots bring a second decision after the first.
No accuracy figure until there is a published result; each report's Appendix A shows how the model matched the seller's own past campaigns.
Security and data
- Your data sits in its own space, separate from other customers'.
- It tunes the model to your products; it never trains a shared model.
- No customer data: sales by product and day, prices and campaign flags are enough.
- Deleted on request; kept no longer than the pilot needs.
Pilot runs are operated by the Jonbar team. No self-serve upload yet.